of Word Pairs Anna-Marie Armstrong School of Psychology, University of Sussex Zoltan Dienes Sussex

Transcription

1 Running head: Subliminal Understanding of Negation Subliminal Understanding of Negation: Unconscious Control by Subliminal Processing of Word Pairs Anna-Marie Armstrong School of Psychology, University of Sussex Zoltan Dienes School of Psychology and Sackler Centre for Consciousness Science, University of Sussex Correspondence concerning this article should be addressed to: Anna-Marie Armstrong, School of Psychology, University of Sussex, Brighton BN1 9QH, UK Phone: , 1

2 Abstract A series of five experiments investigated the extent of subliminal processing of negation. Participants were presented with a subliminal instruction to either pick or not pick an accompanying noun, followed by a choice of two nouns. By employing subjective measures to determine individual thresholds of subliminal priming, the results of these studies indicated that participants were able to identify the correct noun of the pair even when the correct noun was specified by negation. Furthermore, using a grey-scale contrast method of masking, Experiment 5 confirmed that these priming effects were evidenced in the absence of partial awareness, and without the effect being attributed to the retrieval of stimulus-response links established during conscious rehearsal. Keywords: Subliminal priming, Subjective thresholds, Guessing criterion, Contrast masking, Unconscious cognition 2

3 Subliminal Understanding of Negation 1. Introduction Just how much information and knowledge can be acquired through subliminal perception, or just how intelligent unconscious cognitive processing is, remains a familiar and controversial theme (Greenwald, 1992; Norman, 2010). In a classic priming experiment, subjects are briefly presented with a word, or prime, that is prevented from entering conscious perception through the use of a forward or backward mask. When subsequently presented with a further target word, participants are quicker to categorise the target if both the prime and target are semantically related. Whilst the unconscious analysis of letters is more sophisticated than the analysis of individual lines or angles, the semantic analysis of subliminal words or even multiple word-strings would indicate a far more intelligent and sophisticated interpretation of unconscious cognition (Loftus & Klinger, 1992). Evidence suggests that the subliminal presentation of a word facilitates lexical and semantic access (e.g., Abad, Noguera & Ortells, 2003; Carr & Dagenbach, 1990; Dell Acqua & Grainger, 1999; Forster & Davis, 1984; Fowler, Wolford, Slade & Tassinary, 1981; Gaillard et al., 2006; Marcel, 1983; Ortells, Daza & Fox, 2003), although the precise interpretation of these results will be addressed below. Subliminal psychodynamic activation (SPA) studies offer evidence of some of the most sophisticated subliminal priming effects, apparently demonstrating the semantic analysis of multiple word primes (Bronstein & Rodin, 1983; Nissenfeld, 1979; Silverman, Ross, Adler & Lustig, 1978; Silverman & Weinberger, 1985; Waller & Barter, 2005). However, SPA studies have been heavily criticised by others that have tried and failed to replicate results (Allen & Condon, 1982; Condon & Allen, 1980; 3

4 Heilbrun, 1980). Furthermore, whatever the replicability of the results, given that the sentences used differ in the specific words used, any effect evidenced may instead be attributable to simple single-word priming. In fact, there still exists controversy regarding whether or not the semantic analysis of subliminal primes even occurs (Abrams & Greenwald, 2000; Damian, 2001; Hutchison, Neely, Neill & Walker, 2004; Kouider & Dupoux, 2004). In an article investigating the extent of unconscious cognition, Greenwald (1992) argued that unconscious cognitive processing is far less sophisticated in its analytical capabilities than is often reported. Greenwald s (1992) argument rests on the premise that additional research has demonstrated unconscious analysis and processing of nothing more elaborate than word fragments. As an example, Abrams and Greenwald (2000) required participants to categorise a set of consciously perceived parent primes as either positive or negative in valence. Participants were subsequently required to categorise a set of subliminally perceived hulip-type hybrid primes, a non-word hybrid of two positive or two negative parent primes (e.g., humour-tulip-hulip, smut-bile-biut). Results indicated that participants were successfully able to categorise emotional valence despite the nonsensical nature of the hybrid primes. In a follow up study, having consciously categorised parent primes, participants were required to positively or negatively categorise a set of so called tumour-type hybrid primes. These primes were similarly created by combining two congruent parent primes to create a semantically comprehensible prime of different valence to parent primes (e.g. humour-tulip-tumour, smut-bile-smile). Results indicated that participants continued to classify emotional valence according to the valence of the parent prime rather than tumour-type prime, even to the extent that smile was categorised as negative. The results of this study compellingly suggest that words are 4

5 analysed at the level of (consciously primed) word-parts as opposed to whole-word meaning. However, Sklar et al. (2012) have suggested that subliminal processing may have appeared limited in past research because of the small time windows that processing is given for backward masked stimuli (typically in the order of 30 ms). Their solution was to use continuous flash suppression, maintaining stimuli as subliminal for as long as two seconds. Impressively, they found that semantically incoherent sentences (e.g., The bench ate a zebra ) broke through suppression faster than coherent sentences (e.g. The lion ate a zebra ). However, as with SPA studies, there may have been a word-level effect influencing breakthrough as literally different words were used in the different conditions (in this example, bench versus lion ). In a second series of experiments, they found that three-term subtractions (e.g ) (though not additions) primed the speed of pronunciation of the subsequent correct answer. In neither the sentence coherence nor three-term subtraction experiments did the stimuli constitute obvious set phrases that may have been previously well learnt as a unit. These results raise the question of what sort of combinations of stimuli are possible to process subliminally. For example, Van Opstal, Gevers, Osman and Verguts (2010) demonstrated that a same/different judgement task on consciously perceived number targets (e.g., 1-1 or 1-3) extended to subliminal letter stimuli (e.g., a-a or a-d) even when participants were unaware of the presence of the letters. Van Opstal, Calderon, Gevers and Verguts (2011) extended this finding by demonstrating that responding to the subliminal same/different judgements (e.g., a-a) could be modulated by unconscious context (e.g., either a-a or a-d). Therefore, priming effects were dependent upon the processing of both elements. We similarly wished to demonstrate semantic priming of two-element (word) primes and unconscious cognitive control by 5

6 investigating whether it is possible to process instructions to exclude (i.e., negation) subliminally. As we will discuss, negation has a special place in consciousness research. The use of negation allows easy control of stimuli, because stimuli can consist of the same words, just with or without not. The use of negation also addresses one of the theoretical limits assigned to unconscious processes. According to Jacoby, Lindsay, and Toth (1992), what the conscious is uniquely equipped to do is control behaviour by excluding certain responses. Unconscious control exerted by subliminal stimuli was investigated by, for example, Lau and Passingham, (2007), in which a subliminal shape indicated which of two tasks to perform; and by van Gaal, Ridderinkhof, Scholte, and Lamme (2010), in which a subliminal no-go cue slowed down responses and activated a frontal-parietal inhibition network (see van Gaal, de Lange, & Cohen, 2012, for a review of related work). Van den Bussche, Segers, and Reynvoet (2008) indicated limits to unconscious control in that the proportion of conscious stimuli could be used to modulate responding but not the proportion of subliminal stimuli. In contrast to previous studies that have looked at subliminal control, we will be exploring it in the specific case of linguistic negation processing. Although not dealing with linguistic negation, the previous work is encouraging in showing that there exists a mechanism by which unconscious control could operate. In this respect, the current work is consistent with Dienes and Perner s (2007) cold control theory of hypnosis, which postulates that hypnosis consists of unconscious executive control. It is also consistent with the findings of, for example, Norman, Price and Jones (2011) and Wan, Dienes, and Fu (2008), who showed people could exert control over the use of structural knowledge, even when it was unconscious. That is, while the processing of subliminal linguistic negation has not been shown, it is plausible that the unconscious can deal with control and exclusion. Thus, the subliminal processing of negation in two-word phrases presents itself as possible on those 6

7 theories that allow unconscious control (contrast Jacoby et al.), but beyond what has so far been shown to occur subliminally. The present set of studies attempted to assess whether, contrary to Abrams and Greenwald (2000), subliminal perception is sensitive to the semantic comprehension of word combinations and sentence structure. In summing up his argument against complex unconscious cognition, Greenwald (1992) issued an empirical two-word challenge. This two-word challenge asserts that to demonstrate successful subliminal priming of twoword primes, neither word should individually impart the final meaning. Therefore, to claim successful unconscious processing of multiple words, each word would need to be individually processed. The present experiments aimed to meet this challenge by presenting participants with a two word instruction, instructing them which of two subsequent words to choose. Therefore, performance would depend on the successful semantic processing of both words. One explanation to account for the failure of many studies to demonstrate successful subliminal semantic activation of single or multiple word primes may be due to the adherence to strict objective thresholds using backward masking when measuring subliminality. Objective methods of assessing unconscious cognition presume that trial accuracy, beyond what would be expected by chance, indicates conscious knowledge (Seth et al., 2008). However, objective methods of assessing subliminal perception fail to take into account subjectivity; that is, an individual s conscious awareness of accuracy. The two thresholds differ, with unconscious processing occurring below the subjective threshold but limited unconscious processing below the objective. Therefore, the use of objective methods in measuring subliminal perception and unconscious processing have been heavily criticised for testing not just unconscious cognition, but degraded unconscious cognition (Dienes, 2004, 2008; Lau & Passingham, 2006). This indicates 7

8 that to determine the full extent of unconscious processing, it is necessary to use the subjective threshold (compare Masters, Maxwell, & Eves, 2009; contrast Van den Bussche, Van den Noortgate, & Reynvoet, 2009, who found no significant effect of using objective versus subjective thresholds in a meta-analysis of subliminal priming effects 1 ). Therefore, using subjective methods of measuring subliminality, the following series of experiments required the participants to choose between two common nouns (e.g. kite-moon ), having been subliminally instructed with which noun to choose (e.g. pick kite, or not kite ). Correct identification of the instructed noun would then indicate that unconscious cognition is capable of both processing and comprehending more complex demands, such as the pick and not instructions in this study. Whilst it could be argued that success in the pick conditions may not necessarily demonstrate the semantic comprehension of pick but rather simple recognition processes or partial word analysis (e.g., Abrams & Greenwald, 2000), success in the not conditions would require the participant to inhibit initial recognition processes. In turn, this inhibition of recognition processes would imply lexical and semantic comprehension of negation. Therefore, if correct identification is above chance expectations then this would indicate that cognition is capable of processing word combinations outside of conscious perception, as measured by the guessing criterion (Cheesman & Merikle, 1984, 1986) and/or the zero-correlation criterion (ZCC). Like Sklar et al. (2012), we will attempt to determine the limits of subliminal perception when it is given more time to operate than allowed by objective thresholds found with backward masking. 1 The mean effect for subjective thresholds was 0.85 (SE 0.5) and for objective, 0.68 (SE 0.24). While the difference is non-significant, a rough Bayes Factor calculated on the difference (0.17, SE 0.55), using a uniform from 0 to 0.85, is 0.87, indicating the non-significant result is insensitive (as the Bayes factor is between 1/3 and 3), and no conclusions follow from this contrast (see Dienes, 2011, for more on Bayes Factors, which are also explained in more detail below). Note also that these studies were not designed to test the difference between subjective and objective thresholds under otherwise equivalent conditions, unlike, for example, Cheesman and Merikle (1984). 8

9 2. Experiment 1 Current investigations into subliminal perception and unconscious cognition have shown the superior priming effects of practiced versus novel primes (Abrams & Grinspan, 2007; Abrams, Klinger & Greenwald, 2002; Draine & Greenwald, 1998). That is, the priming effects of subliminal primes that have earlier been perceived as conscious targets prove more successful than non-practiced novel primes. This effect has been attributed to consciously perceived primes creating an episodic memory trace which is later re-activated upon subsequent subliminal presentation (Forster & Davis, 1984). Therefore, to achieve maximum likelihood of successful subliminal priming effects, all subliminal primes in Experiment 1 were first practiced as a series of conscious trials. It was expected that for the conscious trials, participants would identify the correct noun in both pick and not conditions on close to 100% of the trials. For the subliminal trials, it was hypothesised that, using a subjective threshold, participants would identify the correct noun for pick and not conditions beyond chance expectations (that is, beyond 50% correct). In addition, the inhibition of recognition processes necessary in not conditions makes it likely that noun identification in pick conditions would be faster than noun identification in not conditions. Therefore, it was hypothesised that response times to noun identification in pick conditions would be faster than in not conditions for both conscious and subliminal trials. In this first experiment we attempted to make the effect likely to occur, so that its absence would be informative. To anticipate, in subsequent experiments we tighten up alternative explanations to determine if the effect goes away Method 9

10 Design & Participants In a repeated measures design with the number of correct identifications being the dependant variable, 25 undergraduate psychology students from the University of Sussex took part in this study in exchange for course credits. Fifteen of the participants were female and ten male, with ages ranging from 18 to 40 years (M = 22.63, SD = 8.52) Apparatus and Materials The experiment was presented on a Dell laptop with a 60Hz screen refresh rate, limiting minimum stimulus presentation to 16 ms, with 16 ms increments. The study was created using E-Prime version 2.0. Trials were created from 20 common nouns, making up a total of 10 noun-pairs (e.g. baby-yard and ant-sky ). All nouns were between 3-5 letters in length, and noun-pairs were phonemically and semantically distinctive and matched in length. Each screen display was centrally presented in lower-case, black, bold Courier New font, and point size 18 on a white background. The arrangement of each of the 10 noun-pairs and instructions were counterbalanced so that participants viewed each of the eight permutations for each noun-pair (e.g. pick yard...1. baby, 2. yard, pick yard...1. yard, 2. baby, not yard...1. baby, 2. yard and not yard...1. yard, 2. baby etc.), creating a total of 80 distinct conscious and subliminal trials Procedure Participants were tested individually in a small quiet space in which they sat before a laptop such that their eyes were approximately 60 cm from the monitor screen. All participants had normal or corrected to normal vision, and English was the first language for all participants. All trials consisted of four separate components: a fixation cross presented for 350ms, the stimulus instruction (e.g. pick baby or not baby ), a backward mask, longer in length than the stimulus and in the form of a series of ampersands (i.e., &&&&&&&) presented for 150ms, and the final component consisting of the noun-pair choice (e.g

11 baby, 2. yard ) in which the participant was required to indicate the number corresponding to the noun in which they had been instructed to choose. The experiment was separated into four continuous phases; conscious trials, SOA setting, subliminal trials, and re-testing the SOA threshold to check for drift Conscious Trials. Having read the instructions, the procedure began with a set of 6 practice conscious trials to accustom the participant to the task required. The common nounpairs used in all practice trials were different from those used in experimental conscious and subliminal phases. Following the fixation cross, the stimulus instruction was presented for 350 ms to ensure conscious perception. Programming in E-Prime ensured that the offset of the stimulus instruction was immediately followed by the onset of the backward mask in all experimental trials. This was especially important for subliminal trials in order to eliminate conscious visual perception. After the backward mask, participants were presented with the noun-pair choice in which they were required to press 1 if they had been instructed to choose the first word, and 2 if they had been instructed to choose the second. The noun-pair choice remained on the screen until the participant had made their choice. Having made their choice, a 250 ms pause preceded the onset of the next trial. Having completed the set of 6 practice trials, participants were instructed to continue to the experimental conscious trials. The procedure for the conscious trials followed the exact procedure used in the practice trials. Participants completed two blocks of 40 randomly presented conscious trials, with an emphasis placed on accuracy as opposed to speed. Participants were not informed whether their choice was correct or incorrect SOA Setting. The SOA of each participant was assessed separately to ascertain individual subjective thresholds. Following the two blocks of conscious trials, participants moved on to the SOA setting phase. Participants were required to complete the same task format used in the conscious phase. Participants were presented with the fixation cross and 11

12 the instruction prime, followed immediately by a backward mask and then the noun-pair choice. Following each trial, participants were required to rate, on a scale of %, how confident they were that they had chosen the correct noun; 100% would indicate that the participant absolutely knew which noun to choose, whilst 50% would indicate that they were purely guessing. During this part of the experiment, if a participant rated confidence to be anything above 50%, stimulus duration was reduced by 16 ms after each trial, from a starting point of 140 ms. Once a participant had rated confidence to be at 50% (guessing), the SOA remained at that same presentation speed for the following trials. Once confidence had been rated at 50% (chance performance) for five successive trials, the experiment proceeded to the subliminal phase. If during any of these five successive trials participants rated confidence to be anything above 50%, SOA was again reduced until five successive trials at 50% confidence had been completed. Before the SOA setting phase began, participants completed a set of 6 practice trials to accustom themselves to the confidence procedure. For the practice trials, prime presentation was held at 140 ms. The common noun-pairs used in both practice and SOA setting phases were different from those used in conscious and subliminal phases Subliminal Trials. Once the SOA setting phase had been completed, the subliminal phase of the experiment consisted of the same 80 trials used in the conscious phase, divided into the same two blocks of 40 randomly placed trials. There were no practice trials for the subliminal phase. Stimulus duration for the subliminal trials was determined by the point at which participants had rated confidence to be at 50% for five successive trials during the SOA setting phase. To prevent rhythmic pressing of the 1 and 2 keys, and to remind participants of the task required, each block of 40 subliminal trials additionally contained 10 randomly placed conscious trials (at 350 ms exposure) (cf. Eimer, Kiss, Press & Sauter, 2009), creating two blocks of 50 trials. 12

13 Threshold Drift. The final phase of the experiment aimed to assess whether individual subjective thresholds of awareness had drifted through the course of the experiment. If the SOA at the finish of the experiment was lower than at the beginning of the subliminal trials, this could indicate that participants may have been consciously aware of the subliminal primes (Kouider & Dupoux, 2004). The SOA threshold drift phase followed the exact format used in the SOA setting phase, using the same materials, with 16 ms decrements in presentation speed from a starting point of 140 ms. Once the participant again rated confidence to be at 50% for five successive trials, the participants were thanked and the experiment ended. After completion of the experiment, participants were fully debriefed and received an information sheet giving some background to the study as well as experimenter details Results SOA Setting. Subjective threshold durations ranged from an SOA of 16 ms to 64 ms, with an average experimental subliminal presentation speed of 48 ms (SD = 15) Trial Accuracy. It was expected that for the conscious phase of the experiment, participants would get approximately 100% of the trials correct. In fact, the mean number of correct identifications for conscious trials was slightly off 100% (M = 97%, SE =.5). For the pick trials, mean correct identification averaged at 95% (SE =.9), whilst for not trials, mean correct identification averaged at 98% (SE =.4). Mean correct noun identification for subliminal trials was 62% (SE = 2), with accuracy for pick (M = 66%, SE = 2) and not (M = 59%, SE = 2) conditions being presented in figure 1, with a 50% reference line indicating chance performance. For all statistical tests, we used an alpha level of.05 to determine significance. Accuracy in both 13

14 pick (t(24) = 7.46, p <.001, d = 3.05) and not (t(24) = 3.9, p =.001, d = 1.59) conditions significantly differed from what would be expected by chance. In addition, a paired-sample t-test looking at the percentage of occasions participants simply chose the subliminally presented noun (i.e. ignoring the preceding instruction) significantly differed between pick (M = 66%, SE = 2) and not (M = 41%, SE = 2, t(24) = 5.97, p <.001, d = 2.44) conditions. Such discrimination was also assessed in terms of (logistic) d, which differed significantly from zero, M =.60, SE =.11, t(24) = 5.62, p <.001, d = Figure 1. Mean percentage values for correct identification of the noun in subliminal pick and not conditions for Experiments 1, 2, 3, 4 and 5 with a 50% reference line Response Time. 14

15 The time taken to identify the instructed noun was recorded for both conscious and subliminal pick and not conditions. For the conscious trials, a paired-sample t-test suggested that on average, participants were significantly quicker to identify the noun in pick conditions (M = 712 ms, SE = 20) than in not conditions (M = 844 ms, SE = 32, t(24) = -5.92, p <.001, d = 2.42). Similarly, for the subliminal trials, a paired-sample t- test suggested that on average, participants were significantly quicker to identify the noun in pick conditions (M = 864 ms, SE = 50) when compared to not conditions (M = 894 ms, SE = 50, t(24) = -2.27, p =.03, d = 0.93) Threshold Drift. Data from the threshold drift phase reveals that subjective threshold durations ranged from an SOA of 32 ms to 80 ms, with a mean experimental subliminal presentation speed of 48 ms (SD = 16.24), matching the sample mean value found in the SOA setting phase, t(24) =.04, p =.97, d = For 15 of the 25 participants, SOA s at the finish of the subliminal trials differed from the SOA at the start of the subliminal phase. Subjective thresholds reduced by 16 ms for seven of the participants, and by 32 ms for one participant. For six of the participants, SOA increased by 16 ms, and for one participant the SOA increased by 48 ms. There was a significant relationship between the SOA setting stage and the SOA threshold drift phase, r =.4, p =.04, indicating there was some consistency in measuring the threshold Trial Accuracy and Response Time. When the data from the eight participants whose SOA had reduced by 16 ms was removed, d significantly differed from zero (M =.56, SE =.15, t(16) = 3.60, p =.002, d = 1.8), and accuracy in both subliminal pick (M = 64%, SE = 3, t(16) = 4.81, p <.001, d = 2.41) and not (M = 58%, SE = 3, t(16) = 2.63, p =.02, d = 1.32) conditions significantly differed from what would be expected by chance. Similarly, when the 15

16 instruction is ignored, the percentage of occasions participants simply chose the subliminally presented noun significantly differed between pick (M = 64%, SE = 3) and not (M = 42%, SE = 3, t(16) = 3.8, p =.002, d = 1.90) conditions. In addition, on removal of the eight participants, participants remained significantly quicker to identify the noun in subliminal pick conditions (M = 845 ms, SE = 64) when compared to not conditions (M = 1139 ms, SE = 86, t(16) = -2.66, p =.02, d = 1.33) Discussion Participants in the subliminal pick condition correctly identified the noun on an average 66% of the trials, whilst correct noun identification in subliminal not trials averaged 58%. Therefore, as hypothesised, participants successfully identified the correct noun at above chance expectations for both subliminally presented pick and not trials. Whilst it could be argued that correct identification in the subliminal pick trials may have demonstrated the ability of unconscious processing to merely recognise letter patterns, correct identification in the subliminal not trials would require the inhibition of these recognition processes. Furthermore, the occasions in which the participant simply chose the subliminally presented noun significantly differed between pick and not conditions, further demonstrating the appropriate use of the subliminal instruction. Consequently, the success of Experiment 1 in demonstrating successful subliminal priming in the not condition may demonstrate the semantic comprehension of not. The use of the guessing criterion for establishing subliminal perception could be criticized on the grounds that participants come with different interpretations as to what guess means. However, in the instructions, and on each screen shot when participants were required to rate confidence, they were given a definition of what guessing (and know ) means. The participants were told to give a value of 50% if they believe that they were purely guessing; that they had no idea which word to choose and that they may as 16

17 well have tossed a coin. They were also told that if they had any confidence at all, if they believed they saw anything of potential relevance at all, they were to give a value above 50. Poorly defined end points on a confidence scale can render the guessing criterion meaningless; thus, the instructions precisely defined the required concept of guess. Further support for the unconscious processing of negation in subliminal conditions was provided by response time data, which demonstrated the difference in cognitive difficulty between pick and not instructions. Once the word pick has been read and cognitively processed, the word indicates that the accompanying noun is the correct noun to choose. Therefore, upon presentation of the noun-pair choice, the letter mapping and recognition processes required to identify the just-presented noun respond quickly. However, the word not indicates that the accompanying noun is not the correct noun to choose. Consequently, upon presentation of the noun-pair choice, it is first necessary to identify the just-presented noun using the same letter mapping and recognition processes used in pick trials before then indicating the other noun. Therefore, the additional time required to indicate the correct noun in not conditions should be evident in both conscious and subliminal response times. Response times for the conscious trials suggested that, as predicted, it took significantly longer to identify the noun in not conditions when compared to pick conditions, an average 131 milliseconds longer. Although it only took an average of 30 milliseconds longer to identify the noun in subliminal not conditions when compared to pick conditions, this difference in response times was also significant, thereby demonstrating the difference in task difficulty, even though participants were not consciously aware of which noun to choose. Past research investigating the extent of subliminal priming paints a controversial and confusing picture. Whilst some studies clearly demonstrate successful (e.g., Diaz & 17

18 McCarthy, 2007; Ortells, Daza & Fox, 2003), and even sophisticated (e.g., Silverman, Ross, Adler & Lustig, 1978; Silverman & Weinberger, 1985) semantic subliminal priming, other studies suggest that the unconscious analysis of words is actually only completed at the sublexical level (e.g. Abrams & Greenwald, 2000; Hutchison, Neely, Neill & Walker, 2004). Experiment 1 aimed to successfully demonstrate the cognitive processing of subliminally presented two-word instructions using individual subjective thresholds. That is, if the individual believed they did not know the correct noun to choose, it can be assumed that they did not possess conscious knowledge (Dienes, 2008). However, whilst the results of Experiment 1 appear to have demonstrated successful unconscious semantic processing, threshold drift data suggests that for eight of the participants, subliminal subjective thresholds may have reduced between SOA settings phases and completion of the subliminal trials. This in turn may indicate conscious, as opposed to unconscious, knowledge of which noun to choose for some of the participants. In addition to potential conscious awareness, significant criticisms arise due to the use of practiced versus novel primes (Damian, 2001; Kunde, Kiesel & Hoffmann, 2003; Schlaghecken & Eimer, 2004). Abrams and Grinspan (2007) argue that simple processing at the feature level is all that is needed to identify a stimulus that is predicted by experience and expectation. As mentioned previously, when primes are practiced consciously they acquire memory traces between a given stimulus and response. These stimulus-response (S-R) mappings remain in short-term memory and are later re-activated upon presentation of the same trials presented subliminally. Whilst these S-R mappings may result in successful subliminal priming, it indicates that the semantic analysis of subliminal primes need not necessarily occur as the semantic system is by-passed. That is, participants may simply have formed an S-R link between, for example, not baby and yard (although yard was associated with each button press 18

19 equally in this situation). A subsequent correct response merely relies on the successful retrieval of the established S-R link and not the semantic processing of not. Experiment 1 used conditions that were most likely to find a priming effect if there were one, and so the results motivate further and more rigorous testing of subliminal priming. Therefore, the issue of practiced versus novel primes and S-R mappings are explored further in Experiment Experiment 2 In Experiment 2, participants performed the same task performed in Experiment 1; a set of conscious trials were followed by an SOA setting phase, a set of subliminal trials and finally a threshold drift phase. However, separate sets of common nouns were used in conscious and subliminal trials to avoid potential successful subliminal priming being attributed to the retrieval of S-R links. To achieve maximum likelihood of successful priming without the establishment of S-R links, participants first practiced pick and not trials consciously with one set of nouns. Participants then consciously viewed the list of nouns that would be used in subliminal trials, in an attempt to activate word representations, before continuing with the experiment. In this way, any positive results could not be credited to the retrieval of S-R links as at no point had the subliminal nouns been paired with any particular response Method Design & Participants In a repeated measures design with the number of correct identifications being the dependant variable, 25 undergraduate psychology students from the University of Sussex took part in this study in exchange for course credits. None of the participants took part in 19

20 Experiment 1. Eighteen of the participants were female and seven male, with ages ranging from 18 to 44 years (M = 20.96, SD = 6.2) Apparatus and Materials Apparatus for Experiment 2 replicated that used in Experiment 1. The 10 noun-pairs used in Experiment 1 were used as conscious trials in Experiment 2, with an additional 20 common nouns between 3-5 letters in length creating a further 10 phonemically and semantically distinctive noun-pairs for subliminal trials Procedure Procedure replicated that of Experiment 1, however following the conscious trials and before the SOA setting, participants were presented with a list of the 20 nouns that would be used in the subliminal trials. Each noun in the list appeared at the centre of the screen for 2000 ms, with a 150 ms pause between each noun. This list of 20 nouns was presented twice Results SOA Setting. Subjective threshold durations ranged from an SOA of 16 ms to 64 ms, with an average experimental subliminal presentation speed of 48 ms (SD = 15) Trial Accuracy. The mean rate of correct identifications made on conscious trials was 95% (SE =.01). Mean correct identifications was 94% (SE =.8) for pick trials, and 97% (SE =.6) for not trials. The mean number of correct noun identifications for the subliminal pick (M = 64%, SE = 2) and not (M = 58%, SE = 2) conditions are presented in figure 1 with a reference line indicating 50% chance performance. On subliminal trials, accuracy on both pick (t(24) = 8.17, p <.001, d = 3.34) and not (t(24) = 4.31, p <.001, d = 1.76) conditions significantly differed from what would be expected by chance. A pairedsample t-test looking at the percentage of occasions participants simply chose the 20

21 subliminally presented noun (i.e. ignoring the preceding instruction) significantly differed between pick (M = 64%, SE = 2) and not (M = 42%, SE = 2, t(24) = 6.67, p <.001, d = 2.72) conditions. Overall subliminal d values also differed significantly from zero (M =.51, SE =.08, t(24) = 6.63, p <.001, d = 2.71) Response Time. The time taken to identify the noun they had been instructed to choose was again recorded for both conscious and subliminal pick and not conditions. A paired-sample t-test revealed that on average, for the conscious trials, participants were significantly quicker to identify the noun in pick conditions (M = 685 ms, SE = 11) than in not conditions (M = 875 ms, SE = 28, t(24) = -6.53, p <.001, d = 2.67). Whilst the results suggested that participants were similarly quicker in subliminal trials to identify the noun in pick (M = 885 ms, SE = 49) conditions when compared to not conditions (M = 952 ms, SE = 33), a paired-sample t-test revealed that this difference in response times was not significant (t(24) = -1.76, p =.09, d = 0.72). However, from the non-significant result we are unable to determine whether this implies that there is evidence for the null hypothesis, that there would be no difference in response times between subliminal pick and not conditions, or that there is no evidence for any conclusion (Dienes, 2011). To do this, we can use a Bayes Factor. Whilst values under 1/3 are substantial evidence in support of the null hypothesis, values over 3 are seen as substantial evidence in support of the experimental hypothesis (Jeffreys, 1963); a Bayes Factor of 1 indicates the evidence is exactly neutral between the two theories. Values between 1/3 and 3 indicate data insensitivity and no conclusions should be drawn. To calculate the Bayes Factor, it is first necessary to specify the likely mean response time difference. The difference in subliminal response times for pick and 21

22 not conditions in Experiment 1 was 30 ms. Thus, a half normal was used with a standard deviation equal to 30 (as per the guidelines in Dienes, 2011, Appendix). The sample mean difference between subliminal pick and not conditions was 67 ms (SE of the difference = 38), leading to a Bayes Factor of B = 2.46, indicating more support for the experimental hypothesis than the null hypothesis (Bayes Factor greater than 1), but also indicating that the data were not sensitive Threshold Drift. Data from the threshold drift phase reveals that subjective SOA durations ranged from an SOA of 16 ms to 80 ms, with an average experimental subliminal presentation speed of 48 ms (SD = 17), matching the mean value found in the SOA setting phase, t(24) =.65, p =.52, d = However, for 17 of the participants, SOA s at the end of the experiment differed from the SOA at the start of the experiment. Subjective thresholds reduced by an average of 16 ms for ten of the participants, and for seven of the participants, SOA increased by 16 ms. There was a significant relationship between the SOA setting stage and the SOA threshold drift phase, r =.67, p <.001, indicating there was some consistency in measuring thresholds Trial Accuracy and Response Time. When the data from the 10 participants whose SOA had reduced by 16 ms was removed, overall d values remained significantly above zero (M =.47, SE =.1, t(14) = 4.64, p <.001, d = 2.48). Accuracy in both subliminal pick (M = 63%, SE = 2, t(14) = 6.03, p <.001, d = 3.22) and not (M = 57%, SE = 3, t(14) = 2.72, p =.02, d = 1.45) conditions significantly differed from what would be expected by chance. Similarly, when the instruction is ignored, the percentage of occasions participants simply chose the subliminally presented noun significantly differed between pick (M = 63%, SE = 2) and not (M = 43%, SE = 3, t(14) = 4.64, p <.001, d = 2.48) conditions. On removal of the 22

23 10 data sets, the difference in response times between subliminal pick (M = 975 ms, SE = 39) and not conditions (M = 1019 ms, SE = 76), remained non-significant (t(14) = -.55, p =.59, d = 0.29) Discussion The accuracy data from the conscious trials in Experiment 2 replicates that found in Experiment 1. For the subliminal trials, participants correctly identified the noun in pick trials at an average rate of 63%, whilst correct identification in subliminal not trials averaged at 57-58%. The results of Experiment 2 replicate those found in Experiment 1 in that the data appears to support the hypothesis that participants would successfully identify the correct noun, above chance performance, for subliminally presented pick and not instructions. Similarly, when the instruction was ignored, the occasions in which the participant simply chose the subliminally presented noun significantly differed between pick and not conditions, providing further evidence to support the appropriate processing of the subliminal instruction. As in Experiment 1, response time data suggested that for conscious trials, participants were significantly quicker to identify the noun in pick conditions when compared to not conditions, by an average 189 ms. Although participants were on average 67 ms quicker to identify the noun in subliminal pick conditions when compared to not conditions, this difference in reaction time was not statistically significant. However, a Bayes Factor indicated insensitive data not strong enough to yet draw conclusions, albeit with more support for the hypothesis of a difference in response times than for the null hypothesis. Experiment 2 aimed to replicate the findings from Experiment 1, whilst avoiding the assumption that successful subliminal priming was a result of the retrieval of S-R links established during conscious rehearsal of stimuli. By consciously viewing the nouns to be used in subliminal trials, presented individually, participants gained the advantage 23

24 of practiced rather than novel primes (Kunde, Kiesel & Hoffmann, 2003), but were prevented from establishing S-R links by viewing the nouns in the absence of either pick or not instructions, supporting research demonstrating that semantic priming can extend to novel and unpractised stimuli (e.g., Naccache & Dehaene, 2001). The results of Experiment 2 appear to support the contention that participants would successfully discriminate between the two nouns at above chance performance in subliminal trials. However, whilst the results of both Experiments 1 and 2 provide support demonstrating successful unconscious processing of logical negation, threshold drift data from both experiments could suggest that conscious processing may be responsible for success in subliminal pick and not conditions. It has been found, for example, that practice with an initially subliminal task can result in participants learning to be conscious, admittedly over considerably more trials than we used (Schwiedrzik, Singer, & Melloni, 2009, 2011). In both Experiments 1 and 2, the threshold drift phase aimed to determine whether individual subjective thresholds of subliminality remained the same at the start and at the end of the subliminal phases of the experiment. If subjective thresholds at the end of the experimental subliminal condition were lower than at the start, it could be argued that participants may have consciously been aware of the stimulus instruction, and thus possessed conscious knowledge as to which noun to choose. While there was not an overall drift down in subjective thresholds, some participants drifted down whilst some drifted up. When the data from those participants whose SOA had drifted down were excluded, the effect remained intact. However, the presence of changes in the assessed thresholds mean that it is possible there existed trials where perception was conscious. The issues regarding conscious awareness and threshold drift was explored further in Experiment 3. 24

25 4. Experiment 3 Experiments 1 and 2 provide evidence that the cognitive unconscious is capable of analysing the syntactic function of subliminally presented pick and not instructions without attributing the priming effect to the retrieval of established S-R links. However, individual visual thresholds may vary from trial to trial as a result of, for example, dark adaption (Holender, 1986). This variation in visual threshold may in turn allow conscious perception of stimuli that is intended to be subliminal. The threshold drift data from both Experiments 1 and 2 demonstrate this possible variance in subjective thresholds as for a number of participants; the measured SOA differed between the start and finish of the subliminal phase. For those participants whose SOA reduced between SOA setting and threshold drift, conscious perception of subliminal primes may be responsible for any successful priming effects. For those participants whose SOA increased between SOA setting and threshold drift, we cannot be sure there was a simple linear increase. Therefore, Experiment 3 aimed to replicate Experiments 1 and 2 by investigating subliminal processing whilst continually assessing subjective thresholds (cf. Marcel, 1983, who also assessed stability of thresholds throughout the priming phase). This was achieved by requiring participants to rate their confidence in selecting the right noun after each trial in the subliminal phase. As in Experiments 1 and 2, it was hypothesised that for the subliminal trials, participants would correctly identify the noun for both pick and not conditions beyond 50% chance expectation. As evidenced in Experiment 1, it was predicted that response times to noun identification in pick conditions would be faster than in not conditions for both conscious and subliminal trials Method 25

26 Design & Participants In a repeated measures design with the number of correct identifications being the dependant variable, 24 undergraduate psychology students from the University of Sussex took part in this study in exchange for course credits. None of the participants took part in Experiments 1 or 2. Nineteen of the participants were female and five male, with ages ranging from 18 to 32 years (M = 20.21, SD = 3.27) Apparatus and Materials Replicated Experiment Procedure Replicated Experiment 2. Participants were also asked to rate their confidence in choosing the correct noun on a scale of %. Participants were required to rate over 50% if they believed they had any awareness of which noun to choose, and to rate 50% if they believed they were guessing Results SOA Setting. Subjective threshold durations ranged from an SOA of 16 ms to 64 ms, with an average experimental subliminal presentation speed of 48 ms (SD = 17) Trial Accuracy. The mean rate of correct identifications made on conscious trials was 98% (SE =.4). Mean correct identifications was 96% (SE =.8) for pick trials, and 99% (SE =.3) for not trials. For the subliminal trials, only those trials in which participants rated confidence to be at 50% (i.e. guessing) were included in the analysis. Of the 80 subliminal trials, the number of trials upon which each participant rated confidence to be above 50% ranged between 0 and 26 trials (M = 7, SD = 7). Mean percentage correct responses for subliminal pick (M = 58%, SE =.6) and not (M = 51%, SE =.6) 26

27 conditions are presented in figure 1, with a 50% reference line indicating chance performance. On subliminal trials, overall d values significantly differed from zero (M =.20, SE =.02, t(23) = 10.57, p <.001, d = 4.40). Accuracy on pick (t(23) = 12.58, p <.001, d = 5.25) trials was significant, whilst not (t(23) = 1.84, p =.07, d = 0.77) trials did not significantly differ from what would be expected by chance. In the previous two experiments, the subliminal not trials produced an effect approximately 8% above baseline. A Bayes Factor, using a half-normal with SD equal to 8%, was B = 2.09, indicated that the data were insensitive, but if anything supported the hypothesis of a subliminal effect. Furthermore, looking at the percentage of occasions participants simply chose the subliminally presented noun (i.e. ignoring the preceding instruction) significantly differed between pick (M = 58%, SE =.6) and not (M = 49%, SE =.6, t(23) = 10.82, p <.001, d = 4.51) conditions. Only if pick and not were differentially processed could there be a significant difference between pick and not trials in the proportion of times the presented word was selected. Conscious knowledge of the subliminal instruction was also assessed using the zero-correlation criterion (ZCC) to establish whether there was a relationship between confidence and accuracy on trials when the participant rated confidence to be above 50%. The difference in accuracy between guess and any confidence was -.54%, which was not significant (t(23) = 1.69, p =.11, d = 0.70). A Bayes Factor was conducted to assess whether the data supported the null hypothesis that there was no relation between confidence and accuracy. Firstly, the range of effect sizes expected if there were conscious knowledge needed to be specified. The maximum slope was determined by the overall accuracy in Experiment 3 (3%) divided by the proportion of confident responses 27

28 (.08) 2. Therefore, the maximum slope = 37.5%. Using a uniform distribution between 0 and 37.5 (sample M = -.54, SE =.31) produced a Bayes Factor of 0.00, providing strong evidence for the null hypothesis that there was no relation between confidence and accuracy 3. The correlation between confidence and accuracy was additionally measured using Type II d. Type II d did not significantly differ from zero (M = -.01, SE =.01, t(23) = -1.69, p =.10, d = A Bayes Factor was conducted to assess whether the Type II data supported the null hypothesis that there was no relation between confidence and accuracy. Given plausible assumptions, Type II d does not exceed Type I (Barrett, Dienes & Seth, in press). Thus, the alternative hypothesis that there existed some relation between confidence and accuracy (i.e., some conscious perception) was modelled as a uniform distribution between 0 and the mean Type I d of.2. The Bayes Factor of 0.03 provided strong support for the null hypothesis and hence the existence of subliminal perception Response Time. The time taken to identify the noun they had been instructed to choose was recorded for both conscious and subliminal pick and not conditions. For the conscious trials, a paired-sample t-test suggested that on average, participants were significantly quicker to identify the noun in pick conditions (M = 728 ms, SE = 23) than in not conditions (M = 851 ms, SE = 35, t(23) = -4.7, p <.001, d = 1.96). Participants were 2 Let X be a weighted average of the performance above baseline when guessing (G) and when confident (C), with the weights being the proportions of each type of response. That is, X = (1 - pc) * G + pc * C. By definition, our measure of confidence accuracy relation, the slope, is C G. This will be maximum when all guessing responses are at baseline, i.e. when G = 0. In this case, slope = C G = C. Also in this case, X = pc * C, with the G term dropping out. Rearranging, C = X/pc. Thus, since maximum slope = C in this case, maximum slope = X/pc. QED. See, for example, Guo et al. (2013) and Li et al. (2013) for the previous use of this method for the zero correlation criterion. 3 Kanai, Walsh and Tseng (2010) offer a subjective discriminability of invisibility (SDI) index to further discriminate between a lack of confidence as a result of either perceptual or attentional blindness. However, due to the lack of trials in which a stimulus was absent, or an appropriate equivalent, we were unable to apply the SDI in this case. For Type II sensitivity, Maniscalco & Lau (2012) show their meta-d measure is superior in principle to Type II d (see also Barret et al, in press, for confirmation with detailed analyses); however, meta-d is more unstable for small N than Type II d in our experience, so we have used the latter. 28

29 similarly quicker in subliminal trials to identify the noun in pick conditions (M = 834 ms, SE = 45) when compared to not conditions (M = 854 ms, SE = 43), however, as evidenced in Experiment 2, a paired sample t-test revealed that this difference in reaction times was not significant (t(23) = -1.63, p =.12, d = 0.68). The mean effect from Experiments 1 and 2 was 48 ms; this was as the standard deviation of a half-normal, as before. With a sample mean difference between subliminal pick and not conditions of 20 ms (SE of the difference = 13), the Bayes Factor was B = 1.47 indicating data insensitivity and no conclusions should be drawn, with the evidence slightly telling against the null hypothesis Discussion Using individual subjective thresholds (Cheesman & Merikle, 1984), the results of Experiments 1 and 2 suggested that when presented with a subliminal prime instruction to choose a particular noun, unconscious cognition is able to successfully choose the correct noun above mere chance performance. Experiment 3 validated the threshold-setting procedure used in the previous experiments. The ZCC indicated a sensitive confirmation of the null hypothesis of no conscious awareness, ruling out partial awareness (Kouider & Dupoux, 2004). Note that partial awareness of the displayed noun in itself is not sufficient to know in any way which choice to make; a participant would need to consciously have partial information to both the noun and the instruction ( not versus pick ). Any such awareness should be reflected in confidence ratings; the ZCC, by contrast, supports the claim that perception was subliminal. It could be argued that maybe participants gave up on using the confidence scale (despite clearly using it appropriately on conscious trials). Even this objection cannot be plausibly sustained because the Bayes factor which indicated strong evidence for the ZCC assumed that the population effect could be indefinitely small. Thus, the alternative hypothesis that was 29

30 rejected is consistent with participants trying to some degree but in a noisy way (i.e. giving up to some degree). The giving up hypothesis, to survive this test, would need to assert a priori that participants gave up completely. Without any prior basis for asserting complete failure to follow instructions, the giving up hypothesis can be rejected. There were trials on which participants indicated some confidence. The results for the ZCC implied that participants used confident responses when they had no better access to information than when they used guess responses. Thus, participants may have been driven by a need to vary the response used, and thus sometimes gave a confidence greater than 50%. Such a tendency will add noise to measuring the threshold, partly explaining the lack of 100% reliability in threshold measurement, and also indicating how the apparent threshold drift in some participants in previous experiments could have been spurious. Because noun pairs were repeated, it might be argued that if a noun pair that had been confidently seen were repeated the subliminal choice effect we observed may in fact depend on consciously primed specific stimulus-response links. However, trials were not repeated exactly, as noun pairs were only repeated for counter-balancing reasons. Thus having once associated a given noun with a left response, there is a higher probability that that same noun will be associated with right response on its next appearance. Thus, S-R links would induce subjects to make incorrect rather than correct responses. Further, the results of the ZCC indicate that confident responses may not have reflected conscious perception, but rather, for example, a desire to use all response options. Whilst an effect of not versus pick remained in Experiment 3, the demonstration would be stronger if the accuracy of not trials were individually significantly above baseline performance. However, research has demonstrated that the 30

31 type of mask used, for example a string of letters or ampersands, can adversely influence the processing of stimuli by interfering with phoneme, grapheme and semantic interpretation (Di Lollo, Enns & Rensink, 2000; McClelland, 1978; Perfetti & Bell, 1991; Walley & Weiden, 1973). Therefore, Experiment 4 aimed to develop a more sensitive method of delivering subliminal stimuli. 5. Experiment 4 To render a prime unconscious, it is necessary to mask the prime in order to avoid conscious perception. The most common method of masking is to use backward masks in the form of symbols (e.g. hatch marks or ampersands), or letter strings (Kiesel, Kunde & Hoffmann, 2007). However, previous research has highlighted the detrimental effect that backward masking has on the cognitive comprehension of subliminal primes (Di Lollo, Enns & Rensink, 2000; McClelland, 1978; Perfetti & Bell, 1991; Walley & Weiden, 1973), due to interference during the pattern and letter recognition part of processing (Grainger, Diependaele, Spinelli, Ferrand & Farioli, 2003). Kouider and Dehaene (2007) state that for a prime to be subliminal, it needs to be presented for a sufficiently short duration, and the mask needs to either share stimulus features or fit the contours of the prime closely. Therefore, Experiment 4 attempted to successfully demonstrate subliminal semantic priming using a grey-scale contrast masking method established by Lamy, Mudrik and Deouell (2008). The experiment followed the same format as Experiment 3 in that participants viewed the list of nouns to be used in subliminal trials to gain the advantage of practice without the establishment of S-R links, and continually assessed subliminal subjective thresholds. As in Experiments 1, 2 and 3, it was hypothesised that for the subliminal trials, participants would correctly identify the noun for both pick and not conditions beyond 31

32 50% chance expectation. As evidenced in the previous 3 experiments, it was predicted that response times to noun identification in pick conditions would be faster than in not conditions for both conscious and subliminal trials. 5.1 Method Design & Participants In a repeated measures design with the number of correct identifications being the dependant variable, 22 undergraduate psychology students from the University of Sussex took part in this study in exchange for course credits. None of the participants took part in Experiments 1, 2, or 3. Sixteen of the participants were female and six male, with ages ranging from 18 to 31 years (M = 20.23, SD = 3.44) Apparatus and Materials Replicated that used in Experiments 2 and Procedure Conscious Trials. The stimulus instruction was presented within a rectangular box of the same size as used for the fixation. The background of the rectangle was filled with grey at a contrast level set by equally altering the red, green and blue (RGB) channels to 212 on the computer monitor. The stimulus instruction was presented within this box in grey at an RGB contrast level of 108 (see figure 2 for an example). The stimulus instruction was presented on the screen for 250 ms to ensure conscious perception. The stimulus instruction was immediately followed by the two-noun choice (e.g. 1. baby and 2. yard ) presented in the centre of the screen SOA Setting. In an attempt to block conscious perception of the prime instruction, each stimulus instruction was presented at an RGB level of 208 against a background RGB contrast of 212 (see figure 3 for an example). 32

33 Figure 2. Example of a conscious contrast contrast mask Figure 3. Example of an unconscious contrast mask Subliminal Trials. The subliminal phase of the experiment contained three blocks of 40 subliminal trials (with the third block being a replication of the first block, at an RGB level of 208 against a background RGB contrast of 212). The presentation time of the stimulus instruction was determined by the point at which the participant rated confidence to be at 50% for five successive trials in the SOA setting phase. Confidence ratings were taken after each trial. Randomly placed within each block of 40 subliminal trials was an additional 10 conscious trials (with the stimulus instruction at an RGB of 108 against a background RGB contrast of 212, presented for 300 ms) to prevent rhythmic pressing of the 1 and 2 keys, and to remind participants of the task required (cf. Eimer, Kiss, Press & Sauter, 2009) Results SOA Setting. Subjective threshold durations ranged from an SOA of 32 ms to 112 ms, with an average experimental subliminal presentation speed of 56 ms (SD = 21) Trial Accuracy. The mean number of correct identifications for conscious trials was slightly off 100% (M = 97%, SE = 1). For the pick trials, mean correct identification averaged at 33